I recently consulted one big telecom and helped to solve their sporadic performance problem which had troubled them for some months. It was an interesting case as it happened in the Oracle / OS touchpoint and it was a product of multiple “root causes”, not just one, an early Oracle mutex design bug and a Unix scheduling issue – that’s why it had been hard to resolve earlier despite multiple SRs opened etc.
Martin Meyer, their lead DBA, posted some info about the problem and technical details, so before going on, you should read his blog entry and read my comments below after this:
So, the problem was, that occasionally the critical application transactions which should have taken very short time in the database (<1s), took 10-15 seconds or even longer and timed out.
- When the problem happened, the CPU usage also jumped up to 100% for the problem duration (from few tens of seconds up to few minutes).
- In AWR snapshots (taken every 20 minutes), the “cursor: pin S” popped into top TOP5 waits (around 5-10% of total instance wait time) and sometimes also “cursor: pin S wait on X” which is a different thing, also “latch: library cache” and interestingly “log file sync”. These waits had then much higher average wait times per wait occurrence than normal (tens or hundreds of milliseconds per wait, on average).
- The V$EVENT_HISTOGRAM view showed lots of cursor: pin S waits taking very long time (over a second, some even 30+ seconds) and this certainly isn’t normal (Martin has these numbers in his blog entry)
AWR and OS CPU usage measurement tools are system-wide tools (as opposed to session-wide tools).
Slides of a presentation about this topic:
I’ll explain the flow of troubleshooting and reasoning here:
As the symptoms involved CPU usage spikes, I first checked whether there were perhaps logon storms going on due a bad application server configuration, where the app server suddenly decides to fire up hundreds of more connections at the same time (that happens quite often, so it’s a usual suspect when troubleshooting such issues). A logon storm can consume lots of CPU as all these new processes need to be started up in OS, they attach to SGA (syscalls, memory pagetable set-up operations) and eventually they need to find & allocate memory from shared pool and initialize session structures. This all takes CPU.However the logons cumulative statistic in AWR didn’t go up almost at all during the 20 minute snapshot, so that ruled out a logon storm. As the number of sessions in the end of AWR snapshot (compared to the beginning of it) did not go down, this ruled out a logoff storm too (which also consumes CPU as now the exiting processes need to release their resources etc).
It’s worth mentioning that log file sync waits also went up by over an order of magnitude (IIRC from 1-2ms to 20-60 ms per wait) during the CPU spikes. However as log file parallel write times didn’t go up so radically, this indicated that the log file sync wait time was wasted somewhere else too – which is very likely going to be CPU scheduling latency (waiting on the CPU runqueue) when CPUs are busy.
As one of the waits which popped up during the problem was cursor: pin S, then I chcecked V$MUTEX_SLEEP_HISTORY and it did not show any specific cursor as a significant contention point (all contention recorded in that sleep history buffer was evenly spread across many different cursors), so that indicated to me that the problem was likely not related to a single cursor related issue (a bug or just too heavy usage of that cursor). Note that this view was not queried during the worst problem time, so there was a chance that some symptoms were not in there anymore (V$MUTEX_SLEEP_HISTORY is a circular buffer of few hundred last mutex sleeps).
So, we had CPU starvation and very long cursor: pin S waits popping up at the same time. cursor: pin S operation should happen really fast as it’s a very simple operation (few tens of instructions only) of marking the cursor’s mutex in-flux so its reference count could be bumped up for a shared mutex get.
Whenever you see CPU starvation (CPUs 100% busy and runqueues are long) and latch or mutex contention, then the CPU starvation should be resolved first, as the contention may just be a symptom of the CPU starvation. The problem is that if you get unlucky and a latch or mutex holder process is preempted and taken off CPU by the scheduler, the latch/mutex holder can’t release the latch before it gets back onto CPU to complete its operation! But OS doesn’t have a clue about this, as latches/mutexes are just Oracle’s memory structures in SGA. So the latch/mutex holder is off CPU and everyone else who gets onto CPU may want to take the same latch/mutex. They can’t get it and spin shortly in hope that the holder releases it in next few microseconds, which isn’t gonna happen in this case, as the latch/mutex holder is still off CPU!
And now comes a big difference between latches and mutexes in Oracle 10.2: When a latch getter can’t get the latch after spinning, it will go to sleep to release the CPU. Even if there are many latch getters in the CPU runqueue before the latch holder, they all spin quickly and end up sleeping again. But when a mutex getter doesn’t get the mutex after spinning, it will not go to sleep!!! It will yield() the CPU instead, which means that it will go to the end of runqueue and try to get back onto CPU as soon as possible. So, mutex getters in 10.2 are much less graceful, they can burn a lot of CPU when the mutex they want is held by someone else for long time.
But so what, if a mutex holder is preempted and taken off CPU by OS scheduler – it should get back onto CPU pretty fast, once it works its way through the CPU runqueue?
Well, yes IF all the processes in the system have the same priority.
This is where a second problem comes into play – Unix process priority decay. When a process eats a lot of CPU (and does little IO / voluntary sleeping) then the OS lowers that processes CPU scheduling priority so that other, less CPU hungry processes would still get their fair share of CPU (especially when coming back from an IO wait for example etc).
When a mutex holder has a lower priority than most other processes and is now taken off CPU, a thing called priority inversion happens. Even though other processes do have higher priority, they can not proceed, as the critical lock or resource they need, is already held by the other process with a lower priority who can’t complete its work as the “high priority” processes keep the CPUs busy.
In case of latches, the problem is not that bad as the latch getters go to sleep until they are posted when the latch is released by the holder process (I’ve written about it here). But the priority inversion takes a crazy turn in case of mutexes – as their getters don’t sleep (not even for a short time) by default, but yield the CPU and try to get back to it immediately and so on until they get the mutex. That can lead to huge CPU runqueue spikes, unresponsive systems and even hangs.
This is why starting from Oracle 11g the mutex getters do sleep instead of just yielding the CPU and also Oracle has backported the fix into Oracle 10.2.0.4, where a patch must be applied and where the _first_spare_parameter will specify the sleep duration in centiseconds.
So, knowing how mutexes worked in 10.2, all the symptoms led me to suspect this priority inversion problem, greatly amplified by how the mutex getters do never sleep by default. And we checked the effective priorities of all Oracle processes in the server, and we hit the jackpot – there was a number of processes with significantly lower priorities than all other processes had. And it takes only one process with low priority to cause all this trouble, just wait until it starts modifying a mutex and is preempted while doing this.
So, in order to fix both of the problems which amplified each other, we had to enable HPUX_SCHED_NOAGE Oracle parameter, to prevent priority decay of the processes and set the _first_spare_parameter to 10, which meant that default mutex sleep time will be 10 centiseconds (which is pretty long time in mutex/latching world, but better than crazily retrying without any sleeping at all). That way no process (the mutex holder) is pushed back and kept away from CPU for long periods of time.
This was not a trivial problem, as it happened in Oracle / OS touchpoint and happened not because a single reason, but as a product of multiple separate reasons, amplifying each other.
There are few interesting, non-technical points here:
- When troubleshooting, don’t let performance tools like AWR (or any other tool!) tell you what your problem is! Your business, your users should tell you what the problem is and the tools should only be used for symptom drilldown (This is what Cary Millsap has been constantly telling us). Note how I mentioned the problem and symptoms separately in the beginning of my post – and the problem was that some business transactions (systemwide) timed out because the database response time was 5-15 seconds!
- The detail and scope of your performance data must have at least as good detail and scope of your performance problem!
In other words, if your problem is measured in few seconds, then your performance data should also be sampled at least every few seconds in order to be fully systematic.The classic issue in this case was that the 20 minute AWR reports still showed IO wait times as main DB time consumers, but that was averaged over 20 minutes. But our problem happened severely and shortly within few seconds in that 20 minutes, so the averaging and aggregation over long period of time did hide the extreme performance issue that happened in a very short time.
Next time when it seems to be impossible to diagnose a problem and if the troubleshooting effort ends up going in circles, then you should ask, “what’s the real problem and who and how is experiencing it” and see if your performance data’s detail and scope matches that problem!
Oh, this is a good point to mention that in addition to my Advanced Oracle Troubleshooting/SQL Tuning seminars I also actually perform advanced Oracle troubleshooting consulting too! I eat mutexes for breakfast ;-)
Update: Srivenu Kadiyala has experienced the same problem and has written about it here.